Mars Recruitment Services — Voice Screening Agent

---

A LiveKit Agents demo that conducts a first-round phone screening for prospective Mars settlers using Simplismart inference. The agent greets the candidate, discusses open mission roles from the catalog, runs two short interview questions (background, motivation for Mars), and closes warmly. The voice loop is intentionally simple — a single Agent with one system prompt, no function tools, no task groups.

After the call ends, the full transcript is sent to an evaluation LLM once to produce a structured JSON evaluation (executive summary, recommendation, score, strengths/concerns, per-question notes). That evaluation is rendered into a rich markdown report at ./reports/interview_<date>_<time>_<sessionId>.md and, if configured, appended as a row to a Google Sheet. The post-call pipeline lives in report_gen.py so the agent file stays focused on the voice loop.

Project layout

interview_agent.py   # the agent (entrypoint + voice loop)
report_gen.py        # post-call transcript → evaluation → markdown + sheets
google_sheets.py     # Google Sheets writer (service-account auth)
mars_jobs.json       # catalog of open Mars mission roles
res/                 # mission-patch image used in reports
reports/             # generated markdown interview reports (gitignored)
env.example          # sample .env — copy to .env.local and fill in

Prerequisites

• Python 3.11+ (Python 3.13 tested)
• LiveKit Agents >=1.5.6
• A LiveKit Cloud project — credentials at https://cloud.livekit.io/projects/p_/settings/keys
• A Simplismart API key for STT / LLM / TTS — https://app.simplismart.ai/settings?tab=2
• (Optional, for Google Sheets output) A Google Cloud project you admin, a service-account JSON key, and gcloud to set them up. See Google Sheets below. The agent writes the local markdown report regardless.

Setup

1. Python virtual env

python -m venv .venv
source .venv/bin/activate
pip install -U pip
pip install -r requirements.txt

2. Environment variables

Copy env.example to .env.local and fill in the required values:

cp env.example .env.local

Required:

SIMPLISMART_API_KEY=eyJhb...
LIVEKIT_URL=wss://<your-project>.livekit.cloud
LIVEKIT_API_KEY=API...
LIVEKIT_API_SECRET=...

Optional — per-service model overrides. Set only the ones you want to change; unset vars fall back to the defaults shown below (from interview_agent.py and report_gen.py).

# Conversational STT / LLM / TTS (Simplismart, by default).
STT_URL=https://api.simplismart.live/predict
STT_MODEL=openai/whisper-large-v3-turbo

LLM_URL=https://api.simplismart.live
LLM_MODEL=meta-llama/Meta-Llama-3.1-8B-Instruct

TTS_URL=https://api.simplismart.live/tts
TTS_MODEL=canopylabs/orpheus-3b-0.1-ft
TTS_VOICE=                 # optional; model-specific

# Post-call evaluation LLM (independent of the conversational LLM).
# Defaults to Simplismart's gpt-oss-120b.
EVAL_LLM_URL=https://api.simplismart.live
EVAL_LLM_MODEL=openai/gpt-oss-120b

# Google Sheets output. Both required to write rows; see the section
# below for setup. The agent always writes local markdown regardless.
GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json
GOOGLE_SHEETS_SPREADSHEET_ID=

TTS plugin is selected from the URL: endpoints ending in /audio/speech go through openai.TTS (OpenAI-compatible); everything else uses simplismart.TTS (native /tts). Swap TTS stacks by flipping TTS_URL.

Running

Dev mode (auto-reload, prints a playground connection URL + token on startup):

python interview_agent.py dev

Console mode (run the agent in the terminal without LiveKit rooms, for quick local testing):

python interview_agent.py console

Production-style start (no reload):

python interview_agent.py start

Connect the test client at https://agents-playground.livekit.io/

Deploying to LiveKit Cloud (optional)

To run this agent on LiveKit Cloud instead of from your laptop, see the deploy quickstart for the full walkthrough. The short version, from this project's root:

In your code change @server.rtc_session() to use explicit dispatch @server.rtc_session(agent_name="mars-recruiter").

1. Install the lk CLI and authenticate it against your Cloud project:

   lk cloud auth

2. Register and deploy the agent. The first run generates livekit.toml and Dockerfile if they aren't already in the repo:

   lk agent create

3. For subsequent updates, push new versions with:

   lk agent deploy

Your local .env.local isn't uploaded automatically — set SIMPLISMART_API_KEY, EVAL_LLM_*, and the Google Sheets pair on the deployed agent via lk agent env (or the Cloud console) before it will run end-to-end. LIVEKIT_URL / LIVEKIT_API_KEY / LIVEKIT_API_SECRET are injected automatically by Cloud.

Google Sheets (optional)

In addition to the local markdown report in ./reports/, each completed interview is appended as a row to a Google Sheet when configured. Authentication uses a Google Cloud service account JSON key, picked up automatically via GOOGLE_APPLICATION_CREDENTIALS. google.auth.default() detects the env var and hands the service-account credentials to gspread.

If Sheets isn't configured, the agent still writes the markdown report locally and logs a warning — only the sheet row is skipped.

One-time setup

You'll need a Google Cloud project you have admin on project to create service accounts. If you don't have a personal one, create one at https://console.cloud.google.com/projectcreate (or reuse an existing personal project).

1. Install and authenticate gcloud

# macOS
brew install --cask google-cloud-sdk
# or follow https://cloud.google.com/sdk/docs/install

gcloud auth login     # opens a browser — sign in as yourself
gcloud projects list  # confirm your project appears
gcloud config set project <your-project-id> 

Do not use the project number here — use the project id (the string-y one in the first column of gcloud projects list).

2. Enable the Sheets + Drive APIs on that project

gcloud services enable sheets.googleapis.com drive.googleapis.com

Or via the console:

• https://console.cloud.google.com/apis/library/sheets.googleapis.com
• https://console.cloud.google.com/apis/library/drive.googleapis.com

Confirm the project selector in the top bar matches your project before clicking ENABLE.

3. Create a service account and download its key

PROJECT_ID=$(gcloud config get-value project)

gcloud iam service-accounts create mars-interview \
  --display-name="Mars Interview Writer"

SA_EMAIL="mars-interview@${PROJECT_ID}.iam.gserviceaccount.com"

mkdir -p ~/.config/gcp
gcloud iam service-accounts keys create ~/.config/gcp/mars-interview-sa.json \
  --iam-account="$SA_EMAIL"

echo "Share your sheet with: $SA_EMAIL"

The key file lives at ~/.config/gcp/mars-interview-sa.json. Treat it like a password — it's not checked in, but don't paste it anywhere.

4. Share the target Google Sheet with the service account

Service accounts act as themselves, so the sheet must explicitly grant them access:

1. Open the sheet you want results written to.
2. Click Share in the top right.
3. Paste the mars-interview@…iam.gserviceaccount.com email from the previous step.
4. Role: Editor.
5. Uncheck "Notify people" (service accounts can't read email).
6. Click Share.

5. Configure this project

Add to .env.local:

GOOGLE_APPLICATION_CREDENTIALS=/Users/<you>/.config/gcp/mars-interview-sa.json
GOOGLE_SHEETS_SPREADSHEET_ID=<the long segment from the sheet URL>

The spreadsheet id is the long alphanumeric part of https://docs.google.com/spreadsheets/d/<THIS_PART>/edit.

6. Verify

.venv/bin/python scripts/check_sheets.py

Expected output: read ok — first tab: Sheet1 (or whatever your first tab is named).

On the next python interview_agent.py dev run you should see Google Sheets service initialized successfully in the logs. A worksheet named Interview Results is created automatically on first write with columns: Timestamp, Session ID, Candidate Name, Score (1-10), Recommendation, Strengths, Areas for Improvement, Summary, Full Conversation.

Troubleshooting

• 403 "Google Sheets API has not been used in project X…" — The Sheets API isn't enabled on the project that owns the service account. Re-run step 2 for that project.
• 403 "The caller does not have permission" — The sheet isn't shared with the service-account email, or it was shared as Viewer instead of Editor. Re-run step 4.
• "serviceusage.services.list" permission denied — You don't have admin on that project (typical for org-managed projects like livekit-cloud-site). Use a personal project instead.
• gcloud is acting as a service account (gcloud auth list shows a *.iam.gserviceaccount.com as ACTIVE) — run gcloud auth login again as your user.

Swapping models

Change one *_URL / *_MODEL / *_VOICE line in .env.local and restart the agent. Each service is independent.

The conversational LLM (LLM_URL + LLM_MODEL) is Simplismart by default, which prioritizes latency.

The post-call evaluation LLM (EVAL_LLM_URL + EVAL_LLM_MODEL) is independent and defaults to Simplismart's openai/gpt-oss-120b at https://api.simplismart.live. Override either env var in .env.local to point at a different OpenAI-compatible endpoint/model. Uses SIMPLISMART_API_KEY. No latency constraint here, so using a bigger model than the conversational one is fine.

Changing one LLM does not affect the other.

LiveKit docs

This project follows the LiveKit Agents conventions. The lk docs CLI and the LiveKit Docs MCP server are the fastest way to look up current API details — see AGENTS.md for install instructions and usage.